← Back to Leaderboard

The AI CUDA Engineer 👷

67_Conv2d_GELU_GlobalAvgPool • warp_optimized_conv2d_gelu_pool_base

Level 2 • Task 67
import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(
    x: torch.Tensor,
    conv_weight: torch.Tensor,
    conv_bias: torch.Tensor,
) -> torch.Tensor:
    """
    Applies convolution, GELU activation, and global average pooling.

    Args:
        x (torch.Tensor): Input tensor of shape (batch_size, in_channels, height, width)
        conv_weight (torch.Tensor): Convolution weight tensor of shape
            (out_channels, in_channels, kernel_size, kernel_size)
        conv_bias (torch.Tensor): Convolution bias tensor of shape (out_channels)

    Returns:
        torch.Tensor: Output tensor of shape (batch_size, out_channels)
    """
    x = F.conv2d(x, conv_weight, bias=conv_bias)
    x = F.gelu(x)
    x = F.adaptive_avg_pool2d(x, 1)
    x = x.squeeze(-1).squeeze(-1)
    return x


class Model(nn.Module):
    """
    Simple model that performs a convolution, applies GELU, and then performs global average pooling.
    """

    def __init__(self, in_channels, out_channels, kernel_size):
        super(Model, self).__init__()
        conv = nn.Conv2d(in_channels, out_channels, kernel_size)
        self.conv_weight = nn.Parameter(conv.weight)
        self.conv_bias = nn.Parameter(conv.bias)

    def forward(self, x, fn=module_fn):
        return fn(x, self.conv_weight, self.conv_bias)


batch_size = 128
in_channels = 3
out_channels = 16
height, width = 32, 32
kernel_size = 3


def get_inputs():
    return [torch.randn(batch_size, in_channels, height, width)]


def get_init_inputs():
    return [in_channels, out_channels, kernel_size]

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a convolution, applies GELU, and then performs global average pooling.
    """
    def __init__(self, in_channels, out_channels, kernel_size):
        super(Model, self).__init__()
        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size)

    def forward(self, x):
        """
        Args:
            x: Input tensor of shape (batch_size, in_channels, height, width)
        Returns:
            Output tensor of shape (batch_size, out_channels)
        """
        x = self.conv(x)
        x = torch.nn.functional.gelu(x)
        x = torch.nn.functional.adaptive_avg_pool2d(x, 1)
        x = x.squeeze(-1).squeeze(-1)
        return x

batch_size = 128
in_channels = 3
out_channels = 16
height, width = 32, 32
kernel_size = 3

def get_inputs():
    return [torch.randn(batch_size, in_channels, height, width)]

def get_init_inputs():
    return [in_channels, out_channels, kernel_size]
Download Evaluation Download PyTorch Download CUDA Download Profiles

Kernel Information

Operation Name 67_Conv2d_GELU_GlobalAvgPool
Level ID 2
Task ID 67
Kernel Name warp_optimized_conv2d_gelu_pool_base
CUDA Speedup (Native) 1.287x
CUDA Speedup (Compile) 1.799x
CUDA Runtime 0.033 ms
PyTorch Runtime (Native) 0.042 ms
PyTorch Runtime (Compile) 0.059 ms
Correct True
Max Diff (vs. Reference) 0.000000
Model azure-gpt-4o-2024-08-06
Temperature 0.50
View Experiment Progress Details

Related Kernels (Level 2, Task 67 • 67_Conv2d_GELU_GlobalAvgPool)

Rank Kernel Name Runtime (ms) Speedup Native Speedup Compile
🥇 optimized_reduction_shared_base_base 0.03 1.63 2.28
🥇 unrolled_fused_conv_gelu_pool_base 0.03 1.63 2.28
🥇 balanced_workload_distribution_base 0.03 1.63 2.28
4 coalesced_fused_conv_gelu_pool_base 0.03 1.52 2.12
5 balanced_workload_fused_conv_gelu_pool_base 0.03 1.47 2.05
5 fused_shared_base 0.03 1.47 2.05
7 optimized_reduction_conv_gelu_pool_base 0.03 1.37 1.91
7 blocksize_fused_conv_gelu_pool_base 0.03 1.37 1.91
7 minimal_sync_conv_gelu_pool_base_base 0.03 1.37 1.91
7 fused_conv_gelu_pool_base 0.03 1.37 1.91
7 modular_fused_conv_gelu_pool_base 0.03 1.37 1.91
7 modular_fused_conv_gelu_pool_base 0.03 1.37 1.91
7 sync_optimized_conv_gelu_pool_base 0.03 1.37 1.91
14 manual_unroll_fused_conv2d_gelu_pool_base 0.03 1.33 1.86
14 streamed_fused_conv_gelu_pool_base 0.03 1.33 1.86
14 unrolled_conv2d_gelu_pool_base 0.03 1.33 1.86
17 warp_optimized_conv2d_gelu_pool_base 0.03 1.29 1.80
18 warp_reduction_conv2d_gelu_pool_base 0.03 1.25 1.75
19 optimized_thread_mapping_conv_gelu_pool_base 0.04 1.06 1.48
19 conv_gelu_avgpool_unroll_edit_1 0.04 1.06 1.48 
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <cmath>

#define KERNEL_SIZE 3

// GELU activation function
__device__ inline float gelu_activate(float x) {
    return 0.5f * x * (1.f + erff(x / 1.41421356f));
}

// Convolution computation with unrolling for a fixed kernel size
template<int ksize>
__device__ inline float compute_conv_unrolled(
    const float* __restrict__ input,
    const float* __restrict__ weight,
    const int in_channels,
    const int in_h,
    const int in_w,
    const int n,
    const int c_out,
    const int out_row,
    const int out_col
) {
    float sum = 0.0f;
    #pragma unroll
    for (int c_in = 0; c_in < in_channels; c_in++) {
        #pragma unroll
        for (int kh = 0; kh < ksize; kh++) {
            #pragma unroll
            for (int kw = 0; kw < ksize; kw++) {
                int in_row = out_row + kh;
                int in_col = out_col + kw;
                int input_idx = ((n * in_channels + c_in) * in_h + in_row) * in_w + in_col;
                int weight_idx = ((c_out * in_channels + c_in) * ksize + kh) * ksize + kw;
                sum += input[input_idx] * weight[weight_idx];
            }
        }
    }
    return sum;
}

// Warp-level reduction using shuffle
__device__ inline float warp_reduce_sum(float val) {
    // Assuming warp size of 32
    for (int offset = 16; offset > 0; offset /= 2)
        val += __shfl_down_sync(0xffffffff, val, offset);
    return val;
}

// Fused kernel: performs convolution, applies bias, GELU activation, and accumulates for global average pooling
// Each block processes one (n, c_out) pair over the conv output (spatial dimensions out_h x out_w).
extern "C" __global__ void conv2d_gelu_pool_kernel(
    const float* __restrict__ input,
    const float* __restrict__ weight,
    const float* __restrict__ bias,
    float* __restrict__ output,
    const int N,
    const int in_channels,
    const int in_h,
    const int in_w,
    const int out_channels,
    const int out_h,
    const int out_w
) {
    // Grid: blockIdx.y -> batch index (n), blockIdx.x -> output channel (c_out)
    const int n = blockIdx.y;
    const int c_out = blockIdx.x;
    
    const int total_pixels = out_h * out_w;
    float local_sum = 0.0f;

    // Each thread processes a subset of the spatial positions in a grid-stride loop
    for (int idx = threadIdx.x; idx < total_pixels; idx += blockDim.x) {
        int row = idx / out_w;
        int col = idx % out_w;
        float conv_val = compute_conv_unrolled<KERNEL_SIZE>(
            input, weight, in_channels, in_h, in_w,
            n, c_out, row, col
        );
        conv_val += bias[c_out];
        conv_val = gelu_activate(conv_val);
        local_sum += conv_val;
    }

    // Perform warp-level reduction
    local_sum = warp_reduce_sum(local_sum);

    // Use shared memory to reduce across warps in the block
    if (threadIdx.x % 32 == 0) {
        atomicAdd(&output[n * out_channels + c_out], local_sum / float(total_pixels));
    }
}

// Host function that launches the fused kernel
torch::Tensor forward(
    torch::Tensor input,
    torch::Tensor conv_weight,
    torch::Tensor conv_bias
) {
    TORCH_CHECK(input.is_cuda(), "input must be a CUDA tensor");
    TORCH_CHECK(conv_weight.is_cuda(), "conv_weight must be a CUDA tensor");
    TORCH_CHECK(conv_bias.is_cuda(), "conv_bias must be a CUDA tensor");

    const int N = input.size(0);
    const int in_channels = input.size(1);
    const int in_h = input.size(2);
    const int in_w = input.size(3);
    const int out_channels = conv_weight.size(0);
    // For a 3x3 filter, output dimensions are (in_h - 2, in_w - 2)
    const int out_h = in_h - 2;
    const int out_w = in_w - 2;

    auto options = torch::TensorOptions().dtype(input.dtype()).device(input.device());
    // Final output has shape (N, out_channels) after global average pooling
    auto final_output = torch::zeros({N, out_channels}, options);

    // Launch the fused kernel with a 2D grid: (c_out, N)
    dim3 grid(out_channels, N);
    const int threads = 256;

    conv2d_gelu_pool_kernel<<<grid, threads>>>(
        input.data_ptr<float>(),
        conv_weight.data_ptr<float>(),
        conv_bias.data_ptr<float>(),
        final_output.data_ptr<float>(),
        N, in_channels, in_h, in_w,
        out_channels, out_h, out_w
    );

    return final_output;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "Warp-optimized Conv2d + GELU + GlobalAvgPool");
}
Performance Metrics
Metric Value Unit Variance Samples
Executed Ipc Active 1.866 inst/cycle 0.000 5
Executed Ipc Elapsed 1.594 inst/cycle 0.001 5
Issue Slots Busy 46.780 % 0.091 5
Issued Ipc Active 1.872 inst/cycle 0.000 5
SM Busy 46.780 % 0.091 5
Memory Throughput 44554574796.030 byte/second 821613541537177856.000 5
Mem Busy 66.226 % 1.825 5
Max Bandwidth 51.134 % 1.079 5
L1/TEX Hit Rate 92.078 % 0.002 5
L2 Hit Rate 84.616 % 2.517 5
Mem Pipes Busy 45.162 % 0.852 5
Warp Cycles Per Issued Instruction 27.518 cycle 0.011 5
Warp Cycles Per Executed Instruction 27.620 cycle 0.010 5
Avg. Active Threads Per Warp 30.670 0.000 5
Avg. Not Predicated Off Threads Per Warp 29.190 0.000 5
Max Active Clusters 0.000 cluster 0.000 5
Max Cluster Size 8.000 block 0.000 5
Overall GPU Occupancy 0.000 % 0.000 5
Cluster Occupancy 0.000 % 0.000 5
Block Limit SM 32.000 block 0.000 5
Block Limit Registers 8.000 block 0.000 5
Block Limit Shared Mem 32.000 block 0.000 5
Block Limit Warps 8.000 block 0.000 5
Theoretical Active Warps per SM 64.000 warp 0.000 5
Theoretical Occupancy 100.000 % 0.000 5
Achieved Occupancy 80.258 % 0.211 5
Achieved Active Warps Per SM 51.362 warp 0.086 5
Analysis Rules
Rule Description
INF HighPipeUtilization ALU is the highest-utilized pipeline (23.3%) based on active cycles, taking into account the rates of its different instructions. It executes integer and logic operations. It is well-utilized, but should not be a bottleneck.
INF CPIStall Check the Warp Stall Sampling (All Cycles) table for the top stall locations in your source based on sampling data. The Kernel Profiling Guide (https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-reference) provides more details on each stall reason.
WRN Occupancy This kernel's theoretical occupancy is not impacted by any block limit. The difference between calculated theoretical (100.0%) and measured achieved occupancy (80.8%) can be the result of warp scheduling overheads or workload imbalances during the kernel execution. Load imbalances can occur between warps within a block as well as across blocks of the same kernel. See the CUDA Best Practices Guide (https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#occupancy) for more details on optimizing occupancy.
Operation / Metric Value Unit
aten::to
CPU Time 380010.12 μs
Device Time 80.51 μs
Self CPU Time 79.78 μs
Self Device Time 0.00 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::zeros
CPU Time 5381517.37 μs
Device Time 188571.28 μs
Self CPU Time 118642.30 μs
Self Device Time 0.00 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::zero_
CPU Time 6653533.58 μs
Device Time 6337113.50 μs
Self CPU Time 243464.45 μs
Self Device Time 0.00 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
aten::fill_
CPU Time 6410071.03 μs
Device Time 6337113.50 μs
Self CPU Time 332463.14 μs
Self Device Time 6337110.81 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
cudaLaunchKernel
CPU Time 6367248.64 μs
Device Time 2518.61 μs
Self CPU Time 6367248.64 μs
Self Device Time 2518.61 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
conv2d_gelu_pool_kernel
CPU Time 0.00 μs
Device Time 2138231.06 μs
Self CPU Time 0.00 μs
Self Device Time 2138231.06 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
cudaEventRecord
CPU Time 173697.76 μs
Device Time 1019327.09 μs
Self CPU Time 173697.76 μs
Self Device Time 1019327.09 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor<int>, at::detail::Array<char*, 1> >(int, at::native::FillFunctor<int>, at::detail::Array<char*, 1>)
CPU Time 0.00 μs
Device Time 6148542.21 μs
Self CPU Time 0.00 μs
Self Device Time 6148542.21 μs
CPU Memory Usage 0 B
Device Memory Usage 0 B
Self CPU Memory Usage 0 B
Self Device Memory Usage 0 B
Status: Completed 
45292 warnings generated when compiling for host.
Suppressed 45323 warnings (45276 in non-user code, 47 NOLINT).
Use -header-filter=.* to display errors from all non-system headers. Use -system-headers to display errors from system headers as well.
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:21:5 bugprone-easily-swappable-parameters
21 | const int n,
| ^~~~~~~~~~~~
22 | const int c_out,
| ~~~~~~~~~~~~~~~~
23 | const int out_row,
| ~~~~~~~~~~~~~~~~~~
24 | const int out_col
| ~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:21:15: note: the first parameter in the range is 'n'
21 | const int n,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:24:15: note: the last parameter in the range is 'out_col'
24 | const int out_col
| ^~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:56:5: warning: 2 adjacent parameters of 'conv2d_gelu_pool_kernel' of similar type ('const float *__restrict') are easily swapped by mistake [bugprone-easily-swappable-parameters]
56 | const float* __restrict__ weight,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
57 | const float* __restrict__ bias,
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:56:31: note: the first parameter in the range is 'weight'
56 | const float* __restrict__ weight,
| ^~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:57:31: note: the last parameter in the range is 'bias'
57 | const float* __restrict__ bias,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:59:5: warning: 2 adjacent parameters of 'conv2d_gelu_pool_kernel' of similar type ('const int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
59 | const int N,
| ^~~~~~~~~~~~
60 | const int in_channels,
| ~~~~~~~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:59:15: note: the first parameter in the range is 'N'
59 | const int N,
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:60:15: note: the last parameter in the range is 'in_channels'
60 | const int in_channels,
| ^~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:62:5: warning: 3 adjacent parameters of 'conv2d_gelu_pool_kernel' of similar type ('const int') are easily swapped by mistake [bugprone-easily-swappable-parameters]
62 | const int in_w,
| ^~~~~~~~~~~~~~~
63 | const int out_channels,
| ~~~~~~~~~~~~~~~~~~~~~~~
64 | const int out_h,
| ~~~~~~~~~~~~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:62:15: note: the first parameter in the range is 'in_w'
62 | const int in_w,
| ^~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:64:15: note: the last parameter in the range is 'out_h'
64 | const int out_h,
| ^~~~~
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:68:19: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
68 | const int n = blockIdx.y;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:69:23: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
69 | const int c_out = blockIdx.x;
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:75:20: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
75 | for (int idx = threadIdx.x; idx < total_pixels; idx += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:75:60: warning: narrowing conversion from 'unsigned int' to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
75 | for (int idx = threadIdx.x; idx < total_pixels; idx += blockDim.x) {
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:98:19: warning: the parameter 'input' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
98 | torch::Tensor input,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:99:19: warning: the parameter 'conv_weight' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
99 | torch::Tensor conv_weight,
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:100:19: warning: the parameter 'conv_bias' is copied for each invocation but only used as a const reference; consider making it a const reference [performance-unnecessary-value-param]
100 | torch::Tensor conv_bias
| ^
| const &
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:106:19: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
106 | const int N = input.size(0);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:107:29: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
107 | const int in_channels = input.size(1);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:108:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
108 | const int in_h = input.size(2);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:109:22: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
109 | const int in_w = input.size(3);
| ^
/home/robert_sakana_ai/llm_cuda/experiments/20250203_optimize_b10_s4_e0_sweep/level_2/task_67/b5_s1_warp_optimized_conv2d_gelu_pool/base/base.cu:110:30: warning: narrowing conversion from 'int64_t' (aka 'long') to signed type 'int' is implementation-defined [bugprone-narrowing-conversions]
110 | const int out_channels = conv_weight.size(0);
| ^